Belangrijkste concepten
SlimSAM introduces a novel data-efficient compression method for Segment Anything Model (SAM) that achieves superior performance with minimal training data.
Samenvatting
Current SAM compression methods require extensive data for training new networks.
SlimSAM reduces training data requirements significantly while maintaining performance.
The alternate slimming framework enhances knowledge inheritance under limited data availability.
Disturbed Taylor pruning addresses misalignment between pruning objectives and training targets.
SlimSAM outperforms existing compression methods with over 10 times less training data.
Statistieken
SlimSAM achieves approaching performance while reducing parameter counts to merely 1.4% (9.1M), MACs to 0.8% (23G), and requiring only 0.1% (10k) of the SAM training data.